This analysis has a lot of moving parts. First, the anonymized mobility data provides counts of visit, counts of visitors, and duration of stay from an origin Census block group to a destination of interest. In our case, the destinations of interest are the 140 points-of-interest that the data vendor has monitored and that correspond to urban green spaces.


Characterizing Census Block Groups

The Richmond Metropolitan Statistical Area consists of:

Cities (FIPS Code) Counties (FIPS Code)
City of Richmond (51760) Amelia County (51007)
City of Petersburg (51730) Caroline County (51033)
City of Hopewell (51670) Charles City County (51036)
City of Colonial Heights (51570) Chesterfield County (51041)
Dinwiddie County (51053)
Dinwiddie County (51053)
Goochland County (51075)
Hanover County (51085)
Henrico County (51087)
King William County (51101)
New Kent County (51127)
Powhatan County (51145)
Prince George County (51149)
Sussex County (51183)

What was the median household income for the Richmond Metropolitan Statistical Area according to the 2019 American Community Survey (ACS) data? We retrieve this value in the code chunk below.

The median household income for Richmond is estimated at $68,529 per the 2015-2019 ACS. Low income block groups were defined as 50% of the area median income ($34,265), while high income block groups were defined as 150% of this value ($102,794).


ggplot(st_as_sf(richmond.msa.bgs.sf)) +
  geom_sf(aes(fill = as.factor(Income.Class))) +
  scale_fill_manual(values = c("grey85", "springgreen4", "darkorange",  "#5B9BD5"), labels = c("Missing Data", "Low Income (50% of median)", "Middle Income", "High Income (150% of median)")) +
  theme(axis.title.x=element_blank(), legend.title = element_blank(), 
        axis.text.y=element_blank(), axis.title.y=element_blank(),axis.text.x=element_blank(),
        axis.ticks = element_blank(), panel.grid.major = element_blank(),
        panel.grid.minor = element_blank(), panel.background = element_rect(fill = "white")) +
  ggtitle(label = "Median Household Income By Census Block Group", 
          subtitle = "Richmond, VA Metropolitan Area") +
  labs(caption="Note: Median household income for the Richmond MSA \n was $68,529 per the 2015-2019 ACS 5-year Estimates.", size = 1.5) +
  theme(plot.caption = element_text(size =8))

ggplot2::ggsave("./Manuscript Images/census_block_groups_by_income_acs_2019.png", width = 12, height = 8, bg = "white")


richmond.msa.bgs.low.income <- richmond.msa.bgs.sf[richmond.msa.bgs.sf$Income.Class == 1, ]
richmond.msa.bgs.mid.income <- richmond.msa.bgs.sf[richmond.msa.bgs.sf$Income.Class == 2, ]
richmond.msa.bgs.hi.income <- richmond.msa.bgs.sf[richmond.msa.bgs.sf$Income.Class == 3, ]


Access to Urban Green Spaces

The location and quality of parks and other urban green spaces are not uniformly distributed within cities and Richmond is no exception. The interactive map below shows the location of the 140 urban green spaces considered shaded according to management type (i.e., who owns and maintains the space) as well as acreage reflected in the size of the point symbol.

parks <- st_read("./Access to UGS/Richmond_POI_Nature_NAICS_712190_Joined_SP_Meters.shp")
## Reading layer `Richmond_POI_Nature_NAICS_712190_Joined_SP_Meters' from data source `C:\Users\bw6xs\Documents\Conferences\UAA\Access to UGS\Richmond_POI_Nature_NAICS_712190_Joined_SP_Meters.shp' 
##   using driver `ESRI Shapefile'
## Simple feature collection with 140 features and 36 fields
## Geometry type: POINT
## Dimension:     XY
## Bounding box:  xmin: 3513940 ymin: 1093382 xmax: 3677249 ymax: 1228239
## Projected CRS: NAD83 / Virginia South
tmap_mode("view")

tm_shape(richmond.msa.bgs.sf, name = "Census Block Groups") +
  tm_polygons(col = "Income.Class", palette = c("grey75", "springgreen4", "darkorange",  "#5B9BD5"), 
          labels = c("Missing Data", "Low Income (50% of median)", "Middle Income", "High Income (150% of median)"), 
          style = "cat", alpha = 0.8, border.col = "white",
              colorNA = "white",
              textNA = "Missing Data", 
              title = "Income Category",
              popup.vars = c("Percent White: " = "PCTWHITE", "Percent African-American: " = "PCTAFAM", 
                             "Percent Hispanic: " = "PCTHISP")) +
tm_shape(parks, name = "Urban Green Spaces") +
  tm_dots(shape = 21, size = "GISACRE", col = "MATYPE", palette = "Spectral", style = "cat", id = "MANAME", 
                    colorNA = "orange", textNA = "Unknown", title = "UGS",
          popup.vars = c("Name: " = "MANAME", "Size (acres): " = "GISACRE" )) +
  tmap_options(check.and.fix = TRUE) + 
  tm_view(set.view = c(-77.494769, 37.513747, 9)) + 
  tm_basemap(c("Esri.WorldGrayCanvas", "CartoDB.Positron", "Esri.WorldTopoMap", "CartoDB.DarkMatter"))


The World Health Organization (WHO)1 has established a protocol for assessing accessibility to parks and green space that considers proximity and size. Rather than relying on a single threshold, this approach estimates the percentage of the population within 300 meters (0.186 miles) of the boundary of a green space at least 1 hectare (2.47 acres) in size and of at least 0.5 hectare (1.23 acres), with the buffer distance explained as “corresponding to [an] approximately 5 minute walk” (p. 32). A third criterion centers on whether there is a green space of at least 100 hectares (~248 acres) within 5000 meters and of the 744 Census block groups in the Richmond region, only 48 meet all three of these WHO criteria. However, 85 percent (i.e., 41 of 48 block groups) fall within the “Middle Income” or “High Income” categories which is slightly lower than the overall distribution of 87 percent (i.e., 650 of 744 block groups) within the “Middle Income” or “High Income” categories.



We also calculate a cumulative measure based on the gravity model (Hansen, 1959) conceptualization of accessibility adopted from Sharifi et al. (2021):

\[A_{i} = \sum_{j=1}^{n} \frac{Size_{j}}{e^{D_{ij}}}\]
In the above equation, A is an accessibility index for Census block group i calculated as the sum of the size of green space j divided by the walking time between the centroid of Census block group i to the boundary of green space j such that more remote destinations contribute less to residents’ green space accessibility, but are still considered. Accounting for accessibility of green space within the study is crucial because prior research has shown that there are significant inequities in access due to the uneven distribution of green space in urban areas.


Of the 186 Census block groups within the upper quartile (i.e., most access) of UGS potential Ai 164 (i.e., 88 percent) fall within the “Middle Income” or “High Income” categories.


Changes in Mobility

Reported cases of COVID-19 increased within the Richmond MSA from March until July of 2020, while hospitalizations peaked in late April.


During the early days of the COVID-19 pandemic people traveled less due in large part to government-mandated lockdowns and the abrupt pivot to online education and work. We use anonymized GPS data collected from mobile devices and provided by SafeGraph to explore general patterns of mobility within the Richmond metropolitan area before and after the March 2020 lockdown was instituted.

It should be noted that the mobility data used for this study were derived from mobile devices through passive sensing, where data are captured without direct effort on the part of users. As a result, the proportion of residents in a given Census block group who are actively contributing data varies over time. For context, total coverage for the entire Richmond region in March of 2018 was 135,768 mobile devices for an estimated total population of 1,293,876 residents or 10.5 percent. In March of 2019 coverage increased to about 11.0 percent and fell back to 10.5 percent in March of 2020. The image below shows the number of devices contributing data as a percentage of total population of each Census block group on the Monday immediately after the March 12th lockdown began:



This reduction in mobility is apparent if we examine the median percentage of time mobile devices were observed at home on the Monday two weeks before the March 12th lockdown began:



Versus the Monday immediately after the March 12th lockdown began:



Compared with the Monday two weeks after the March 12th lockdown began:



We can also see that the amount of time spent at home after the lockdown was imposed was mediated by income and race with mobile devices based in Low Income Census block groups spending much less time at home than their counterparts in Middle Income and High Income Census block groups. It is also clear that mobile devices based in Census blocks with a higher percentage of residents identifying as White only tended to spend more time at home during this period.


Changes in UGS Usage

Outdoor activities like visiting parks and other urban green spaces were among the only options for active and passive recreation that continued to function and did not involve coming into close physical contact with other people who might be a source of infection. We can ask How did the number of visits to UGS change before and after the March 2020 lockdown was instituted? and explore this question using the anonymized GPS data collected from mobile devices and provided by SafeGraph. In these datasets, a mobile device must spend at least 4 minutes at a given point of interest (POI) in order to count as a visit.

The pair of barcharts below show that the number of visitors to UGS in the analysis sample declined considerably for mobile devices based in Low Income Census block groups between the first and last weeks of March 2020. Similarly, the grouped barchart shows the number of visitors to UGS in the analysis sample as a percentage—this is an important distinction given that mobility declines and the number of visitors to UGS dropped across the board during the lockdown period.



Neither the UGS access potential metric nor the WHO criteria not appear to be related to these changes in UGS usage and so the statistics and visualizations exploring these relationships are omitted here.



We might expect that Census block groups with greater access to UGS (i.e., physical proximity and/or lower travel time) may have made more visits or stayed longer at these destinations and we can explore this question using the anonymized GPS data collected from mobile devices and provided by SafeGraph. Because the available data only provide a median dwell time in minutes, these are visualized as a series of scatterplots showing each week in March and April of 2020 alongside the corresponding week in March and April of 2019. The SafeGraph documentation define the median dwell time data element as follows:

This is the median of the minimum dwell times we have calculated for each of the visits to the POI (point of interest). We determine the minimum dwell time by looking at the first and last ping we see from a device during a visit. This is a minimum dwell because it is possible the device was at the POI longer than the time of the last ping.

What I see in the plots below is that with the pandemic lockdown, the mobile devices contributing data to SafeGraph frequented UGS in the analysis sample closer to their home Census block group and by the end of April 2020, were staying at the UGS destinations longer. Of course there are other determinants at work including differences in weather.

















Survey of Selected Census Block Groups

We surveyed households in four Census block groups located within the City of Richmond boundaries as a means of controlling—in part—for differential physical access to UGS given that the City of Richmond has been lauded for access to parks and open space afforded its residents. Further, each of the surveyed Census block groups fell within the Low Income category but were selected to ensure variation in the racial and ethnic characteristics of residents. These four block groups also vary in regard to physical access to UGS and one of the four (i.e., 517600413002) meets all of the access criteria established by the WHO. The pop-up menu for each of the four surveyed block groups also contains the names of the nearest UGS and the estimated walk time from the centroid of the Census block group.

The survey itself was constructed using Qualtrics and administered by a professional survey firm called Dynata. The household survey response rate was lower than expected, despite nominal compensation in the form of a $10 gift card as a financial incentive (i.e., received 172 responses from 1,644 invitations for a response rate of 10 percent).

variable n min max mean sd
PCTWHITE 4 19.115 76.917 40.481 27.463
PCTAFAM 4 14.792 76.723 54.894 28.834
PCTASIAN 4 0.000 5.750 1.569 2.798
PCTHISP 4 0.708 6.333 3.723 2.350
TOTPOP 4 762.000 2400.000 1294.250 752.869
UGS_Potential 4 1.487 1.773 1.594 0.131


Survey Findings

In addition to descriptive statistics (not presented here), we used non-parametric Kruskal-Wallis tests to explore potential differences in outcomes interest.


Life Satisfaction

  • How satisfied were you with life during the pandemic? (0 = Not At All, 10 = Completely)
    • Education (N = 154, p = 0.006, effsize = moderate)
    • Ethnicity (N = 153, p = 0.037, effsize = small)
    • Location (N = 154, p = 0.039, effsize = small)
    • Age (N = 153, p = 0.041, effsize = large)
    • Social with friends, family, colleagues before pandemic (N = 152, p = 0.033, effsize = small)
    • Strong sense of belonging before pandemic (N = 150, p = 0.003, effsize = moderate)

Of the 27 respondents who indicated that their life satisfaction increased during the pandemic, 24 (88%) live within walking distance (15-20 minutes) of UGS. Further, 22 of these 27 respondents (81%) spent at least 30 minutes on average in the green space they visited most frequently.


Happiness

  • Overall, how happy did you feel?
    • Education (N = 155, p = 0.014, effsize = moderate)
    • Location (N = 155, p = 0.036, effsize = small)
    • Social with friends, family, colleagues before pandemic (N = 154, p = 0.006, effsize = moderate)
    • Change in community spirit—1 = Much Less, 5 = Much More (N = 155, p = 0.008, effsize = moderate)
    • Minutes spent exercising (N = 147, p = 0.042, effsize = small)
    • Employment status (N = 153, p = 0.034, effsize = moderate)


Stress

  • Did you feel stress during the pandemic? (1 = Not At All, 5 = Very Much)
    • Strong sense of community before pandemic (N = 154, p = 0.037, effsize = small)
    • Noticed change in bird/wildlife presence (N = 153, p = 0.036, effsize = small)
    • Noticed increase in flora (N = 153, p = 0.004, effsize = moderate)


Physical Activity

  • Compared to people of your age, how physically active were you during the pandemic? (1 = Much Less, 5 = Much More)
    • Change in the frequency and /or duration of time spent outdoors—1 = Much Less, 5 = Much More (N = 153, p = 0.000, effsize = moderate)
    • Satisfaction with quality of this green space—1 = Very Dissatisfied, 5 = Very Satisfied (N = 154, p = 0.038, effsize = small)
    • Gender (N = 154, p = 0.015, effsize = small)


Explaining Change in Mental and Physical Well Being

Because the majority of the survey responses were collected on a Likert-type scale, one approach would be to use ordinal logistic regression. This specification takes into account the fact that outcome being predicted (e.g., change in life satisfaction before and during the pandemic) is not continuous and it is also not categorical.

However, we instead use structural equations modeling (SEM) to further explore the relationships between neighborhood-scale environmental perceptions, the likelihood of green space use, and mental wellbeing. We know that access to green spaces offer numerous mental and physical health benefits (Hartig et al. 2014; Kondo et al. 2018) and prior studies have highlighted the importance of this access during the pandemic (Meyer et al., 2020, Cheng et al. 2021; Larson et al. 2022). The rationale has been to test hypothesized relationships between UGS access, physical activity, and mental wellbeing (i.e., the restorative effects of nature as well as active and passive recreation as a coping mechanism) using the 172 survey responses as a means of complementing the existing mobility analysis of the SafeGraph data and the initial statistical analysis of the raw survey responses. Because environmental quality matters as a determinant of green space usage, we retain measures of environmental quality at the neighborhood scale. Although we do not have enough survey responses^[[Kline (2016) recommends that the sample size should be 10 times (or ideally 20 times) as many cases as parameters, and at least 200.] to test the full conceptual framing (i.e., all of the latent variables of interest), we were able to fit a reduced model which suggests that the well-being benefits of greenspace increased during the pandemic.

SEM is an umbrella term that includes path analysis and confirmatory factor analysis. Path analysis is confirmatory factor analysis (CFA) that does not include any latent variables—it is the structural component of SEM. CFA is the measurement model of an SEM that outlines the relationships between indicators (e.g., responses to a questionnaire) and latent variables.

By default, the factor loading of the first indicator of a latent variable is fixed to 1, thereby fixing the scale of the latent variable. The chi-squared test indicates the difference between observed and expected covariance matrices. Values closer to zero indicate a better fit; smaller difference between expected and observed covariance matrices. The RMSEA ranges from 0 to 1, with smaller values indicating better model fit. An RMSEA value of 0.07 or less is indicative of acceptable model fit. The standardized root mean square residual ranges from 0 to 1, with a value of 0.08 or less being indicative of an acceptable model. Based on the literature, the CFI and TLI cut-off scores should be above 0.9 otherwise, we do not have good internal validity.

A commonly used rule of thumb is that an alpha of 0.7 indicates acceptable reliability, however a recent article by Flora (2020) argues that survey responses on the familiar Likert scale should be treated as categorical variable when computing measures of reliability in a SEM context.



First, we need to establish which indicators (i.e., survey questions) adequately measure the latent variables of interest. Typically at least four indicators are needed to achieve this, then we look specifically at measures of reliability—how consistently do the indicators measure the construct of interest? Suggested thresholds for common measures of reliability like alpha and omega range from 0.7 to 0.8.

The Chi-squared test (\(\\X^{2}\)) indicates the difference between observed and expected covariance matrices with the null hypothesis that the data align with (corroborate) the proposed model. Root Mean Square Error of Approximation (RMSEA) is a measure that attempts to correct the tendency of chi-square statistics to reject models with large samples. RMSEA is considered very good if it is equal to or less than 0.05 and good if between 0.05 and 0.08.


Latent Variable: Social Wellbeing

To begin, we leverage selected items from the household survey responses to measure a latent variable social_wb and because the ultimate goal is to fit an SEM (and determine how latent variables like social wellbeing and green space access are related), we assume the survey items are related through a a single common factor (i.e., one factor CFA). We measure can social_wb as follows:


Indicator Mean Standard Dev. Factor Loading
How satisfied were you with life during the pandemic? 5.97 2.20 0.96
How happy were you during the pandemic? 5.67 2.30 0.84
How worthwhile was your life during the pandemic? 6.29 2.48 0.84
UCLA-3 Loneliness score during pandemic 5.88 2.00 0.41


The Alpha (\(\alpha\)) reliability measure is 0.84 and the Omega Hierarchical (\(\omega_{h}\)) reliability measure is 0.86, both of which suggest an acceptable level of internal consistency for this construct. In terms of model fit:

  • The degrees of freedom are 2 and the fit is 0
  • The number of observations was 153 with Chi Square = 0.26 with prob < 0.88 (values of 5 or less are a common benchmark and a non-significant result indicates good fit)
  • RMSEA index = 0 and the 10 % confidence intervals are 0 0.08 (values below 0.08 are typically considered acceptable)


Latent Variable: UGS Access and Benefits

Now, the latent variable ugs_benefits is measured as follows, but most if the indicators were recoded such that 2 = More Beneficial During the Pandemic, 1 = Same Benefit, and 0 = Less Beneficial or N/A:


Indicator Mean Standard Dev. Factor Loading
Does using public green space reduce stress more during the pandemic? 1.33 0.72 0.81
Does using public green space allow you to relax in nature more during the pandemic? 1.31 0.68 0.81
Does using public green space allow you to keep fit more during the pandemic? 1.11 0.78 0.66
Does using public green space allow you to follow a pedestrian route more during the pandemic? 0.76 0.73 0.47
Does using public green space allow you to enjoy wildlife more during the pandemic? 1.12 0.70 0.59
Does using public green space allow you to have fun more during the pandemic? 1.18 0.70 0.64
Does using public green space allow you to get fresh air more during the pandemic? 1.31 0.65 0.78
How often do you visit the UGS you visit most frequently during the pandemic? 3.47 1.84 0.37


The Alpha (\(\alpha\)) reliability measure is 0.85 and the Omega Hierarchical (\(\omega_{h}\)) reliability measure are 0.86, which suggests an acceptable level of internal consistency for this construct. In terms of model fit:

  • The degrees of freedom for just the general factor are 20 and the fit is 0.25
  • The number of observations was 147 with Chi Square = 35.7 with prob < 0.017 (values of 5 or less are a common benchmark and a non-significant result indicates good fit)
  • RMSEA index = 0.073 and the 10 % confidence intervals are 0.031 0.112 (values below 0.08 are typically considered acceptable)



SEM Findings

Takeaways

  1. World Health Organization—WHO. (2016). Urban green spaces and health: A review of evidence. Copenhagen, Denmark: WHO Regional Office for Europe.↩︎